File size: 9,504 Bytes
c8ddb9b |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 |
"""Helper functions for models."""
import pathlib
import pickle
from copy import deepcopy
from pathlib import Path
from typing import Any, List, Dict
import matplotlib.pyplot as plt
import numpy as np
import torch
from PIL import Image
from torch import optim
from src.models.modules.discriminator import Discriminator
from src.models.modules.generator import Generator
from src.models.modules.image_encoder import InceptionEncoder
from src.models.modules.text_encoder import TextEncoder
# pylint: disable=too-many-arguments
# pylint: disable=too-many-locals
def copy_gen_params(generator: Generator) -> Any:
"""
Function to copy the parameters of the generator
"""
params = deepcopy(list(p.data for p in generator.parameters()))
return params
def define_optimizers(
generator: Generator,
discriminator: Discriminator,
image_encoder: InceptionEncoder,
text_encoder: TextEncoder,
lr_config: Dict[str, float],
) -> Any:
"""
Function to define the optimizers for the generator and discriminator
:param generator: Generator model
:param image_encoder: Image encoder model
:param text_encoder: Text encoder model
:param discriminator: Discriminator model
:param lr_config: Dictionary containing the learning rates for the optimizers
"""
img_encoder_lr = lr_config["img_encoder_lr"]
text_encoder_lr = lr_config["text_encoder_lr"]
gen_lr = lr_config["gen_lr"]
disc_lr = lr_config["disc_lr"]
optimizer_g = optim.Adam(
[{"params": generator.parameters()}],
lr=gen_lr,
betas=(0.5, 0.999),
)
optimizer_d = optim.Adam(
[{"params": discriminator.parameters()}],
lr=disc_lr,
betas=(0.5, 0.999),
)
optimizer_text_encoder = optim.Adam(text_encoder.parameters(), lr=text_encoder_lr)
optimizer_image_encoder = optim.Adam(image_encoder.parameters(), lr=img_encoder_lr)
return optimizer_g, optimizer_d, optimizer_text_encoder, optimizer_image_encoder
def prepare_labels(batch_size: int, max_seq_len: int, device: torch.device) -> Any:
"""
Function to prepare the labels for the discriminator and generator.
"""
real_labels = torch.FloatTensor(batch_size, 1).fill_(1).to(device)
fake_labels = torch.FloatTensor(batch_size, 1).fill_(0).to(device)
match_labels = torch.LongTensor(range(batch_size)).to(device)
fake_word_labels = torch.FloatTensor(batch_size, max_seq_len).fill_(0).to(device)
return real_labels, fake_labels, match_labels, fake_word_labels
def load_params(generator: Generator, new_params: Any) -> Any:
"""
Function to load new parameters to the generator
"""
for param, new_p in zip(generator.parameters(), new_params):
param.data.copy_(new_p)
def get_image_arr(image_tensor: torch.Tensor) -> Any:
"""
Function to convert a tensor to an image array.
:param image_tensor: Tensor containing the image (shape: (batch_size, channels, height, width))
"""
image = image_tensor.cpu().detach().numpy()
image = (image + 1) * (255 / 2.0)
image = np.transpose(image, (0, 2, 3, 1)) # (B,C,H,W) -> (B,H,W,C)
image = image.astype(np.uint8)
return image # (B,H,W,C)
def get_captions(captions: torch.Tensor, ix2word: Dict[int, str]) -> Any:
"""
Function to convert a tensor to a list of captions.
:param captions: Tensor containing the captions (shape: (batch_size, max_seq_len))
:param ix2word: Dictionary mapping indices to words
"""
captions = captions.cpu().detach().numpy()
captions = [[ix2word[ix] for ix in cap if ix != 0] for cap in captions] # type: ignore
return captions
def save_model(
generator: Generator,
discriminator: Discriminator,
image_encoder: InceptionEncoder,
text_encoder: TextEncoder,
epoch: int,
output_dir: pathlib.PosixPath,
) -> None:
"""
Function to save the model.
:param generator: Generator model
:param discriminator: Discriminator model
:param image_encoder: Image encoder model
:param text_encoder: Text encoder model
:param params: Parameters of the generator
:param epoch: Epoch number
:param output_dir: Output directory
"""
output_path = output_dir / "weights/"
Path(output_path / "generator").mkdir(parents=True, exist_ok=True)
torch.save(
generator.state_dict(), output_path / f"generator/generator_epoch_{epoch}.pth"
)
Path(output_path / "discriminator").mkdir(parents=True, exist_ok=True)
torch.save(
discriminator.state_dict(),
output_path / f"discriminator/discriminator_epoch_{epoch}.pth",
)
Path(output_path / "image_encoder").mkdir(parents=True, exist_ok=True)
torch.save(
image_encoder.state_dict(),
output_path / f"image_encoder/image_encoder_epoch_{epoch}.pth",
)
Path(output_path / "text_encoder").mkdir(parents=True, exist_ok=True)
torch.save(
text_encoder.state_dict(),
output_path / f"text_encoder/text_encoder_epoch_{epoch}.pth",
)
print(f"Model saved at epoch {epoch}.")
def save_image_and_caption(
fake_img_tensor: torch.Tensor,
img_tensor: torch.Tensor,
captions: torch.Tensor,
ix2word: Dict[int, str],
batch_idx: int,
epoch: int,
output_dir: pathlib.PosixPath,
) -> None:
"""
Function to save an image and its corresponding caption.
:param fake_img_tensor: Tensor containing the generated image
(shape: (batch_size, channels, height, width))
:param img_tensor: Tensor containing the image
(shape: (batch_size, channels, height, width))
:param captions: Tensor containing the captions
(shape: (batch_size, max_seq_len))
:param ix2word: Dictionary mapping indices to words
:param batch_idx: Batch index
:param epoch: Epoch number
:param output_dir: Output directory
"""
output_path = output_dir
output_path_text = output_dir
capt_list = get_captions(captions, ix2word)
img_arr = get_image_arr(img_tensor)
fake_img_arr = get_image_arr(fake_img_tensor)
for i in range(img_arr.shape[0]):
img = Image.fromarray(img_arr[i])
fake_img = Image.fromarray(fake_img_arr[i])
fake_img_path = (
output_path / f"generated/{epoch}_epochs/{batch_idx}_batch/{i+1}.png"
)
img_path = output_path / f"real/{epoch}_epochs/{batch_idx}_batch/{i+1}.png"
text_path = (
output_path_text / f"text/{epoch}_epochs/{batch_idx}_batch/captions.txt"
)
Path(fake_img_path).parent.mkdir(parents=True, exist_ok=True)
Path(img_path).parent.mkdir(parents=True, exist_ok=True)
Path(text_path).parent.mkdir(parents=True, exist_ok=True)
fake_img.save(fake_img_path)
img.save(img_path)
with open(text_path, "a", encoding="utf-8") as txt_file:
text_str = str(i + 1) + ": " + " ".join(capt_list[i])
txt_file.write(text_str)
txt_file.write("\n")
def save_plot(
gen_loss: List[float],
disc_loss: List[float],
epoch: int,
batch_idx: int,
output_dir: pathlib.PosixPath,
) -> None:
"""
Function to save the plot of the loss.
:param gen_loss: List of generator losses
:param disc_loss: List of discriminator losses
:param epoch: Epoch number
:param batch_idx: Batch index
:param output_dir: Output directory
"""
pickle_path = output_dir / "losses/"
output_path = output_dir / "plots" / f"{epoch}_epochs/{batch_idx}_batch/"
Path(output_path).mkdir(parents=True, exist_ok=True)
Path(pickle_path).mkdir(parents=True, exist_ok=True)
with open(pickle_path / "gen_loss.pkl", "wb") as pickl_file:
pickle.dump(gen_loss, pickl_file)
with open(pickle_path / "disc_loss.pkl", "wb") as pickl_file:
pickle.dump(disc_loss, pickl_file)
plt.style.use("fivethirtyeight")
plt.figure(figsize=(24, 12))
plt.plot(gen_loss, label="Generator Loss")
plt.plot(disc_loss, label="Discriminator Loss")
plt.xlabel("No of Iterations")
plt.ylabel("Loss")
plt.legend()
plt.savefig(output_path / "loss.png", bbox_inches="tight")
plt.clf()
plt.close()
def load_model(
generator: Generator,
discriminator: Discriminator,
image_encoder: InceptionEncoder,
text_encoder: TextEncoder,
output_dir: pathlib.Path,
device: torch.device
) -> None:
"""
Function to load the model.
:param generator: Generator model
:param discriminator: Discriminator model
:param image_encoder: Image encoder model
:param text_encoder: Text encoder model
:param output_dir: Output directory
:param device: device to map the location of weights
"""
if (output_dir / "generator.pth").exists():
generator.load_state_dict(torch.load(output_dir / "generator.pth", map_location=device))
print("Generator loaded.")
if (output_dir / "discriminator.pth").exists():
discriminator.load_state_dict(torch.load(output_dir / "discriminator.pth", map_location=device))
print("Discriminator loaded.")
if (output_dir / "image_encoder.pth").exists():
image_encoder.load_state_dict(torch.load(output_dir / "image_encoder.pth", map_location=device))
print("Image Encoder loaded.")
if (output_dir / "text_encoder.pth").exists():
text_encoder.load_state_dict(torch.load(output_dir / "text_encoder.pth", map_location=device))
print("Text Encoder loaded.")
|